Chapter 5 (Eigenvalues and Eigenvectors): Diagonalization (对角化)

本文为《Linear algebra and its applications》的读书笔记

Diagonalization

• The factorization $\mathbf{A}=\mathbf{P}\mathbf{D}{\mathbf{P}}^{-1}$, where $D$ is a diagonal matrix, is used to compute powers of $A$, decouple dynamical systems in Sections 5.6 and 5.7, and study symmetric matrices and quadratic forms in Chapter 7.

---

• Powers of a diagonal matrix are easy to compute. So if $A=PD{P}^{-1}$ for some invertible $P$ and diagonal $D$, then $A^k$ is also easy to compute.
  • For example, if $D=\left[\begin{array}{cc}5& 0\\ 0& 3\end{array}\right]$, then $D^k=\left[\begin{array}{cc}{5}^k& 0\\ 0& 3^k\end{array}\right]$
• A square matrix $A$ is said to be diagonalizable (可对角化) if $A$ is similar to a diagonal matrix, that is, if $A=PD{P}^{-1}$ for some invertible matrix $P$ and some diagonal matrix $D$.

The Diagonalization Theorem

• In other words, $A$ is diagonalizable if and only if there are enough eigenvectors to form a basis of ${\mathbb{R}}^n$. We call such a basis an eigenvector basis of ${\mathbb{R}}^n$.
• 注意，可对角化不一定可逆 (有可能有 0 特征值)

Matrices Whose Eigenvalues Are Distinct

Proof: Check Theorem 2 in Section 5.1

---

EXAMPLE 1

Compute $A^8$, where $A=\left[\begin{array}{cc}4& -3\\ 2& -1\end{array}\right]$.

SOLUTION

• $\det (A-\lambda I)={\lambda }^2-3\lambda +2=(\lambda -2)(\lambda -1)$. The eigenvalues are 2 and 1, and the corresponding eigenvectors are ${\mathbf{v}}_1=\left[\begin{array}{l}3\\ 2\end{array}\right]$ and ${\mathbf{v}}_2=\left[\begin{array}{l}1\\ 1\end{array}\right]$.
• Next, form
  $$P=\left[\begin{array}{ll}3& 1\\ 2& 1\end{array}\right],D=\left[\begin{array}{ll}2& 0\\ 0& 1\end{array}\right],\text{ and }{P}^{-1}=\left[\begin{array}{rr}1& -1\\ -2& 3\end{array}\right]$$
• Since $A=PD{P}^{-1}$,
  $$\begin{array}{rl}{A}^8=P{D}^8{P}^{-1}& =\left[\begin{array}{ll}3& 1\\ 2& 1\end{array}\right]\left[\begin{array}{rr}{2}^8& 0\\ 0& 1^8\end{array}\right]\left[\begin{array}{rr}1& -1\\ -2& 3\end{array}\right]\\ & =\left[\begin{array}{ll}3& 1\\ 2& 1\end{array}\right]\left[\begin{array}{rr}256& 0\\ 0& 1\end{array}\right]\left[\begin{array}{rr}1& -1\\ -2& 3\end{array}\right]\\ & =\left[\begin{array}{ll}766& -765\\ 510& -509\end{array}\right]\end{array}$$

Matrices Whose Eigenvalues Are Not Distinct

PROOF

• a. Suppose that the multiplicity of the eigenvalue ${\lambda }_k$ is $m$, then $det(A-\lambda I)$ has the form $(\lambda -{\lambda }_k{)}^m\cdot ...$, which means that $A-\lambda I$ can be row reduced to a triangular matrix that has $m$ $\lambda -{\lambda }_k$ entries on its main diagonal. Thus when $\lambda ={\lambda }_k$, $A-\lambda I=A-{\lambda }_{k}I$ has at most $m$ non-pivot columns. It indicates that $dimNul(A-{\lambda }_{k}I)$(the dimension of the eigenspace) is at most $m$.
• c. Let { v 1 , . . . , v s } \{\boldsymbol v_1,..., \boldsymbol v_s\} {v1​,...,vs​} be the eigenvectors in the sets { B 1 , . . . , B k } \{\mathcal B_1,...,\mathcal B_k\} {B1​,...,Bk​}. Suppose { v 1 , . . . , v s } \{\boldsymbol v_1,..., \boldsymbol v_s\} {v1​,...,vs​} is linearly dependent. Since ${\mathbf{v}}_1$ is nonzero( { v 1 } \{\boldsymbol v_1\} {v1​} is linear independent), let $r$ be the least index such that ${\mathbf{v}}_1,...,{\mathbf{v}}_r$ are linearly dependent. Then there exist scalars $c_1,...,c_r$ such that
  $$c_1{\mathbf{v}}_1+...+c_r{\mathbf{v}}_r=0(1)$$where $c_1,...,c_r$ are not all 0. Suppose { v p + 1 , . . . , v r } \{\boldsymbol v_{p+1},..., \boldsymbol v_r\} {vp+1​,...,vr​} are eigenvectors corresponding to the same eigenvalue ${\lambda }_{p+1}$. Since a linear combination of { v p + 1 , . . . , v r } \{\boldsymbol v_{p+1},..., \boldsymbol v_r\} {vp+1​,...,vr​} is still an eigenvector corresponding to ${\lambda }_{p+1}$, the equatioin (1) can be transformed into
  $$c_1{\mathbf{v}}_1+...+c_p{\mathbf{v}}_p={\mathbf{w}}_{p+1}(2)$$where $-{\mathbf{w}}_{p+1}=c_{p+1}{\mathbf{v}}_{p+1}+...+c_r{\mathbf{v}}_r$ and ${\mathbf{w}}_{p+1}$ is an eigenvector corresponding to ${\lambda }_{p+1}$. Multiplying both sides of (2) by $A$, we obtain
  $$\begin{gathered} c_{1}A{\mathbf{v}}_1+...+c_{p}A{\mathbf{v}}_p=A{\mathbf{w}}_{p+1} \\ {c}_1{\lambda }_1{\mathbf{v}}_1+...+c_p{\lambda }_p{\mathbf{v}}_p={\lambda }_{p+1}{\mathbf{w}}_{p+1}(3) \end{gathered}$$Multiplying both sides of (2) by ${\lambda }_{p+1}$ and subtracting the result from (3), we have
  $$c_1({\lambda }_1-{\lambda }_{p+1}){\mathbf{v}}_1+...+c_p({\lambda }_p-{\lambda }_{p+1}){\mathbf{v}}_p=0(7)$$Since { v 1 , . . . , v p } \{\boldsymbol v_1,..., \boldsymbol v_p\} {v1​,...,vp​} is linearly independent, the weights in (7) are all zero, which is impossible. Hence { v 1 , . . . , v r } \{\boldsymbol v_1,..., \boldsymbol v_r\} {v1​,...,vr​} cannot be linearly dependent and therefore must be linearly independent.